抄録
As population decline progresses, navigation technology for sensor-equipped robots has advanced, leading to the practical implementation of delivery robots such as Uber’s. Currently, many robots adopt an on-demand detection approach using onboard sensors, selecting routes based on the real-time positions of pedestrians. However, challenges remain, including collision risks and inefficient avoidance behavior. This study aims to improve route selection by introducing a deep learning-based technique to predict pedestrian positions a few seconds ahead. While previous research has demonstrated high-accuracy predictions for bidirectional pedestrian movement, prediction accuracy in complex scenarios remains unclear. If accurate predictions can be made, robots will be able to anticipate pedestrian movements and avoid collisions in advance. Therefore, this study evaluates the effectiveness of predictive detection through the cooperation of spatial AI and agent AI. Specifically, surveillance camera footage is analyzed using YOLO and DeepSort. In on-demand detection, real-time pedestrian positions are identified, and avoidance areas for robots are set. In predictive detection, pedestrian position data is treated as a time series and input into the deep learning model ConvLSTM to predict locations two seconds ahead. The prediction accuracy is evaluated based on the coverage rate, which measures the consistency between the predicted area and the actual pedestrian position. This study aims to achieve more efficient robot navigation through these evaluations.
https://ror.org/0197nmd03 
